Considering that the difficulty of fault feature extraction of rolling bearing vibration signals and the difficulty of automatic and accurate identification of fault categories, a method based on improved ensemble deep auto-encoder(IEDAE) is proposed. Firstly, the loss function of the auto-encoder is improved and three kinds of wavelet convolution auto-encoders are designed. Then, five kinds of auto-encoders, such as discriminative auto-encoder and wavelet convolution auto-encoder, are employed to construct the corresponding deep auto-encoders, and a "cross-layer" connection is designed to alleviate the gradient disappearance of the deep network, and unsupervised pre-training and supervised fine-tuning of bearing vibration signals are realized. Finally, the recognition result is given using the weighted averaging method to ensure accurate and stable diagnosis result. The experimental results show that the IEDAE can effectively identify the bearing faults under multiple working conditions and multiple fault severities, which can get rid of the dependence on manual feature extraction and has better ability of feature extraction and recognition than other existing methods.